RanDepict

RanDepict generates diverse chemical structure depictions to create varied datasets for training deep learning-based optical chemical structure recognition (OCSR) systems.

Key Features:

• Multi-library rendering: Uses depiction functionalities from RDKit, CDK, Indigo, and PIKAChU to produce chemical structure images.
• Diverse depiction parameters: Controls parameters such as bond length, line thickness, and label font style to vary depiction appearance.
• Image augmentation capabilities: Adds features such as curved arrows, chemical labels around structures, and geometric distortions to increase image variability.
• Depiction feature fingerprints: Encodes depiction and augmentation characteristics as binary vectors summarizing image features.
• MaxMin algorithm for diversity: Applies the MaxMin algorithm to select diverse samples from valid depiction and augmentation options.

Scientific Applications:

• OCSR dataset generation: Produces diverse and augmented chemical structure images for training optical chemical structure recognition systems.
• Model robustness and benchmarking: Supplies varied depiction styles and augmentations for testing and benchmarking deep learning-based chemical structure recognition models.

Methodology:

Leverages depiction functions from RDKit, CDK, Indigo, and PIKAChU; varies depiction parameters and applies image augmentations; represents images with binary depiction feature fingerprints and selects diverse samples using the MaxMin algorithm.